KR20240043924A - Method for Allocating Downlink Power in Multi-Cell Multi-User Massive MIMO System - Google Patents

Abstract

The present invention is a multi-cell multi-user Massive MIMO system that optimizes and allocates downlink power to a user (UE) based on the location information of the user (UE) in the multi-cell multi-user Massive MIMO system. It concerns a method of allocating downlink power.
The downlink power allocation method in a multi-cell, multi-user Massive MIMO system according to the present invention is a deep neural network (DNN) model in a Massive MIMO (Multi input multi output) system with multi-cells and multi-users. Based on the location of the user (UE), learning is performed to optimize and allocate the user's (UE) downlink power, and based on the learning results, the optimized downlink power allocation is predicted according to the user's (UE) location. By being used for power allocation, it is provided to reduce computational complexity, improve spectral efficiency and energy efficiency, and allocate downlink power to users (UEs) in real time.

Description

{Method for Allocating Downlink Power in Multi-Cell Multi-User Massive MIMO System}

The present invention relates to a downlink power allocation method in a MIMO system. In particular, in a multi-cell multi-user Massive MIMO system, the downlink power of a user (UE) is optimized and allocated based on the user's (UE) location information. This relates to a downlink power allocation method in a multi-cell, multi-user Massive MIMO system that enables

In a MIMO (multi input multi output) system to transmit large amounts of data at high speed, a base station (BS) uses a large antenna array (hundreds of antennas) to support many users (UEs). The Massive MIMO system not only increases the capacity of the system, but also significantly increases spectral efficiency (SE) and energy efficiency (EE). Therefore, Massive MIMO technology provides high spectral and energy efficiency improvements. It is emerging as an advanced wireless network technology that can provide.

Meanwhile, MISO (multiple-input single-output) NOMA (Non-orthogonal multiple access) is a network technology to support services to multiple users simultaneously. It divides/assigns the same frequency to multiple users with different powers at the same time. It is a technology that supports services to multiple users. In this MISO-NOMA network system, unlike the existing orthogonal multiple access, non-orthogonal multiple access is used to transmit slots (frequency, time, etc.) to multiple users at the same time by allocating different transmission power, so that various users can be connected with one transmission. Data can be transmitted to others. And SWIPT (Simultaneously Wireless Information and Power Transfer) is a technology that transmits information and energy simultaneously, using some of the received signals for energy storage and the rest for data transmission. Recently, research on HUP (Hybrid User Pairing) is being conducted to improve spectral efficiency (SE) and energy efficiency (EE) in the MISO NOMA SWIPT DL (DownLink) system. So far, research has been conducted using existing methods. The study was conducted as a way to solve the UE pairing problem.

Additionally, a method using joint pilot and data fusion was studied to solve the spectral efficiency problem in a multicell environment with active users (UEs). Deep neural network (DNN) methods are being studied for channel prediction in massive MIMO systems, and deep neural network (DNN) methods are also being studied to improve physical layer security (PLS) performance.

However, these prior studies do not consider the problem of allocating power to users (UEs). In other words, it takes a lot of time to allocate power to each user (UE) within the base station (BS), making it inefficient in terms of spectral efficiency and energy efficiency, and because many base stations are deployed to support users (UEs), Although many interference problems occur in a cell multi-user environment, previous studies have not sufficiently studied the problem of allocating power to users (UEs).

Republic of Korea Patent Publication No. 10-1325969 (registered on October 31, 2013) Republic of Korea Patent Publication No. 10-1612663 (registered on April 7, 2016)

The present invention was proposed to solve the power allocation problem of users (UEs) occurring in the conventional multi-cell, multi-user Massive MIMO system. The purpose of the present invention is to solve the problem of power allocation of users (UEs) in the multi-cell, multi-user Massive MIMO system. The purpose is to provide a downlink power allocation method in a multi-cell multi-user Massive MIMO system that can be supported in real time by optimizing the downlink (DL) power allocation of users (UE) based on location information.

The downlink power allocation method in a multi-cell, multi-user Massive MIMO system according to the present invention to achieve the above object is a deep neural network in a Massive MIMO (Multi input multi output) system with multi-cells and multi-users. Through the (DNN) model, learning is performed to optimize and allocate the user's (UE) downlink power based on the user's (UE) location, and based on the learning results, optimized downlink power is optimized according to the user's (UE) location. Predict link power allocation and use it for power allocation.

Here, the deep neural network model calculates SINR (Signal-to-Interference-plus-Noise Ratio) and SLNR (Signal-to-Leakage-plus-Noise Ratio) according to the location of the user (UE), and the calculated SINR Based on and SLNR, Max-Min or Max-Product for power allocation is determined to predict the user's (UE) optimized power allocation.

In the Massive MIMO system, when there are multiple (J) base stations (BS) with multiple (M) antennas in each cell, the SINR and SLNR of user (UE) i in cell l are calculated using the following equation: It is calculated.

(here, is the user channel vector of user (UE) i in cell l, is the precoding vector of user (UE) i in cell l, is the precoding vector of user (UE) k in cell l, is the amount of noise change, and the SLNR (Signal-to-Leakage-plus-Noise-Ratio) per user is (indicated as)

(here, is the user channel vector of user (UE) k in cell l)

Accordingly, the downlink (DL) channel capacity of user (UE) i in cell l for the SINR and SLNR standards is calculated using the following equation.

(here, is the prelog factor for the sample portion per coherence block used for downlink (DL) data)

항등행렬(identity matrix)을 나타낸다)을 통해 산출된다.Meanwhile, for optimal downlink power allocation according to the location of the user (UE), the precoding vector of the user (UE) is selected based on M-MMSE (multicel minimum mean-squared error) to optimize the downlink spectrum. Supports efficiency, but the precoding vector of the user (UE) is (here, is the precoding vector of user (UE) i in cell l, is a fused vector for detecting the uplink signal sent by user (UE) i in cell l) that satisfies Select and is MRT (maximum ratio transmission) fusion Assuming that it is made by The multicel minimum mean-squared error (M-MMSE) fusion of (here, is the channel estimation of user (UE) i in cell l based on MMSE at base station j, as is the transmission power of user (UE) i in cell l, is the estimated error correlation matrix for the channel between base station j and user (UE) i in cell l, is the noise variation of the uplink (UL), In cell ㅣ

It is calculated through the identity matrix.

The max-min objective function and conditions for optimizing the downlink power allocation of the user (UE) are obtained through the following equation.

(here, represents the maximum downlink transmission power)

And the Max-Product objective function and conditions for optimizing the SINR and SLNR of the user (UE) are obtained through the following equation.

To solve the fairness problem that occurs in the Max-Min determined above, the Monte Carlo technique is applied.

Meanwhile, the deep neural network model performs learning to optimize and allocate downlink power through a forward propagation process and a back propagation process using learning data and result data. In the back propagation process, the learning results derived from the forward propagation process are compared with the resulting data using a loss function, and the connection weight between each unit of the deep neural network model is adjusted in the direction of reducing errors. By updating the and bias using the Adam (adpative moment estimation) algorithm, learning is performed to optimize and allocate downlink power.

In addition, the learned deep neural network model is mounted on a server provided in the Massive MIMO system, and the locations of new users (UEs) are generated in each cell, and new users (UEs) are generated from the base station (BS) to which the cell belongs. When the location information of the user (UE) is received, the optimized downlink power allocation according to the location of the user (UE) is predicted using the received location information of the user (UE) as input data, and the predicted power allocation information is sent to the corresponding base station. It is transmitted to (BS) so that it can be used for power allocation of users (UEe) at the base station (BS).

According to the present invention, in a multi-cell multi-user Massive MIMO system, optimized downlink (DL) power of a user (UE) is allocated through a deep neural network (DNN) model based on the user's (UE) location information. By doing so, it has the effect of reducing computational complexity, improving spectral efficiency and energy efficiency, and allocating downlink power to users (UEs) in real time.

1 is an overall conceptual diagram of a multi-cell multi-user Massive MIMO system according to the present invention;
Figure 2 is an example of a power allocation optimization algorithm for solving the Max-Min fairness problem according to the present invention;
Figure 3 is an example of a power allocation optimization algorithm for solving the Max-Product problem according to the present invention;
Figure 4 is a deep neural network (DNN) structure diagram for power allocation in the multi-cell multi-user Massive MIMO system according to the present invention;
Figure 5 is an example of a proposed deep neural network model for power allocation according to the present invention;
Figure 6 is an example of a graph showing system capacity as a function of spectral efficiency (SE) per user (UE) by transmission power when the number of users (UE) is greater than the number of antennas in the base station according to the present invention.
Figure 7 is an example graph showing processing time as a function of number of user (UE) locations in accordance with the present invention.
Figure 8 is an example graph showing CDF as a function of spectral efficiency (SE) per user (UE) in Max-Product downlink according to the present invention;
Figure 9 is an example of a graph showing energy efficiency (EE) as a function of total spectral efficiency according to the present invention.

In the present invention, in order to obtain an optimized solution for power allocation for the downlink (DL) of a multi-cell multi-user Massive MIMO system, computational complexity is reduced based on user (UE) location information, and spectral efficiency and energy efficiency are achieved. improves and proposes a method to support power allocation in real time.

For this purpose, the present invention uses the Signal-to-Interference-plus-Noise Ratio (SINR) and Signal-to-Leakage-plus-Noise Ratio (SLNR) criteria in the Massive MIMO system, Solve the allocation problem. In addition, to support optimal spectral efficiency, a convergence technology of M-MMSE (multicell-minimum mean square error) and precoding of the Massive MIMO system is used.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings.

First, the abbreviations used in the present invention are as shown in Table 1 below.

Figure 1 shows an overall conceptual diagram of a multi-cell multi-user Massive MIMO system according to an embodiment of the present invention.

As shown in Figure 1, the multi-cell multi-user Massive MIMO system model according to the present invention is a deep learning multi-cell multi-system consisting of J base stations (BAs) with M antennas in each cell. A user MIMO system is considered.

Each cell knows and calculates the location information of the user (User Equipment) and sends the results to the server. The server uses the location information of the user (UE) to model the deep neural network model (DNN Model). ) to learn through. The results learned from the deep neural network model are sent to each base station (BS), and are used for power allocation in the Massive MIMO system.

In the present invention, power is allocated to maximize the channel capacity and SINR (or SLNR) of the Massive MIMO system. To maximize the channel capacity, Max-Min is used, and SINR (Signal-to-Interference-plus-Noise Ratio) is allocated. ) and Max-Product are used to maximize SLNR (Signal-to-Leakage-plus-Noise Ratio). Max-Min and Max-Product are widely known methods among various power allocation methods. Max-Min optimizes the performance of the entire system by maximizing the channel capacity of the user with the lowest channel capacity (Spectral Efficiency). Max-Product is a method of maximizing the overall user channel efficiency (SINR or SLNR) by optimizing (maximizing) the product of users' SINR (or SLNR).

In the present invention, SINR (Signal-to-Interference-plus-Noise Ratio) and SLNR (Signal-to-Leakage-plus-Noise Ratio) criteria are used to determine Max-Min and Max-Product in power allocation of the Massive MIMO system. Use . Max-Min and Max-Product are widely known methods among various power allocation methods. Max-Min optimizes the performance of the entire system by maximizing the channel capacity of the user with the lowest channel capacity (Spectral Efficiency). Max-Product is a method of maximizing the overall user channel efficiency (SINR or SLNR) by optimizing (maximizing) the product of users' SINR (or SLNR).

Meanwhile, SLNR reflects the loss capacity of a user (UE) with respect to other users, with high SLNR indicating large channel gain and low loss capacity. At this time, the power loss of a certain user (UE) is valuable because it indicates intrusion from other users (UEs). Therefore, when all users (UEs) have low loss energy, each user (UE) has minimal interference energy with other users (UEs).

And in the case of SINR, a high SINR is advantageous in that it directly indicates a high transmission rate. On the other hand, the calculation of SINR poses a very challenging problem in that it requires knowing information about all other users.

The SINR and SLNR are obtained through the following equations 1 to 3.

First, in Figure 1, the received signal of user (UE) i in cell l is expressed as Equation 1 below.

here, is the user channel vector of user (UE) i in cell l, is additive white Gaussian, is the precoding vector (unitome beamforming vector) of user (UE) i in cell l, represents the downlink (DL) data signal of user (UE) i in cell l, respectively. is the amount of noise change, and the SLNR (Signal-to-Leakage-plus-Noise-Ratio) per user is It is expressed as , and SINR is defined as in Equation 2 below.

here, is the precoding vector of user (UE) k in cell l.

The calculation of SINR in Equation 2 above is for all users. Information is required, and obtaining this information accurately is a very challenging problem.

에 있는 유저

의 SLNR은 다음 수학식 3과 같이 정의된다.Next, cell

user in

The SLNR of is defined as follows in Equation 3.

here, is the user channel vector of user (UE) k in cell l.

과 비교하여

의 측정은 다른 유저 프리코딩 벡터 를 포함하지 않는다. The SLNR represents the power loss of user (UE) i with respect to all other users (UES). Large channel gain and low power are associated with high SLNR. As a result, all users (UE) have low loss capacity, which means that each user (UE) has minimal interference with other users (UEs). It is very important to recognize that the power loss of one user (UE) is an intrusion from the perspective of other users (UEs).

compared to

Measurement of different user precoding vectors does not include

In order to allocate optimized downlink power according to the location of the user (UE), a channel estimation process is necessary.

에 의해서 사용된다. 유저(UE) 당

의 전체 업로드(UL) 파일롯 전력을 사용하여, MMSE(Minimum Mean Square Error)는 다른 방법들과 비교하여 최적으로 채널을 추정할 수 있기 때문에 본 발명에서는 MMSE의 표준 추정 기술을 사용한다. Channel estimation plays a very important role in precoder design because channel estimation with high accuracy is required to obtain the best precoder design. In order to determine channel estimation, the present invention considers the channel vector estimated by pilot-based channel preparation at base station j. It is assumed that the base station and users (UEs) are completely synchronized and operate based on time-division duplex (TDD). Here, the downlink (DL: downlink) data transmission process is performed in the channel estimation learning step in the uplink (UL: uplink). A pilot is present and in every cell

It is used by. per user (UE)

Since Minimum Mean Square Error (MMSE) can estimate the channel optimally compared to other methods by using the total upload (UL) pilot power of , the standard estimation technique of MMSE is used in the present invention.

Therefore, MMSE-based channel estimation between the j-th base station (BS) and the i-th user of the l-th cell ( ) is equivalent to the following equation 4:

은

항등행렬(identity matrix)이다. here is the spatial correlation matrix, is the reciprocal of the channel estimation correlation matrix between base station j and user (UE) i in cell l.

silver

It is an identity matrix.

The pilot signal received according to the progress is expressed in Equation 5 below.

here, represents noise. is the noise change amount, is a circularly symmetric complex Gaussian distribution with zero-mean, and is the estimated error correlation matrix for the channel between base station j and user (UE) i in cell l.

The estimation error is and are independent of each other.

In cell j, the base station (BS) transmits a downlink (DL) signal to the user (UE), which can be expressed mathematically as Equation 6 below.

here, is the downlink data signal for user (UE) i in cell j. Downlink signal precoding vector is assigned to The precoding vector defines the transmission spatial directionality, satisfies, represents the transmission power. In Massive MIMO systems, the capacity limit (referred to as the hardening bound) can be used to calculate the possible downlink spectral efficiency (SE). The ergodic channel capacity of user (UE) i in cell l for the SINR standard is expressed in Equation 7 below.

here is the prelog factor for the sample portion per coherence block used for downlink data.

Similar to Equation 7 above, the capacity for the SLNR standard can be expressed as Equation 8 below.

Here, the channel realization is used to calculate the expected value. The fraction of samples per coherence block used for downlink data is the pre-log factor. It is indicated by . The significance is that the user (UE) can obtain a low bound by using the average value of the precoded channel, and it is a reasonable assumption that shows channel capacity (hardening).

Unlike uplink (UL), downlink (DL) spectral efficiency (SE) is Finding an optimal precoder is challenging because it depends on the precoding vector. Based on duality uplink-downlink, using an appropriate heuristic approach this Select the precoding vector. here, is the precoding vector of user (UE) i in cell l, is a fused vector for detecting the uplink signal sent by user (UE) i in cell l. In the present invention It is assumed that the method of using the transmission antennas of all base stations (BS) uses the MRT (maximum ratio transmission) method, in which all antennas participate in data transmission to improve performance. thus It is assumed that it is created by . here, is the MMSE channel estimation, and is number i of the lth cell in the jth base station (BS)?? M-MMSE (multicel minimum mean-squared error) convergence of channel efficiency between users can be expressed as Equation 9 below.

here, Based on Equation 9, M-MMSE shows optimal results but has high computational complexity, while MRT does not show optimal results but shows the simplest complexity for convergence of methods.

Energy efficiency (EE) is the sum of the spectral efficiency of all users (UE) in one cell divided by the total power consumption in one cell, and can be expressed as Equation 10 below.

유저(UE)에 의해서 소모된 하드웨어 정적 전력(hardware static power consumption)이며, 는 전체 기지국 하드웨어 정적 전력 소모(total BS hardware static power consumption)를 각각 나타낸다.Here, v is the transmission power efficiency, Is

It is hardware static power consumption consumed by the user (UE), represents the total BS hardware static power consumption, respectively.

The downlink spectral efficiency (Equation 7 and Equation 8) of user (UE) i in cell l is averaged to realize small-scale fading, and the downlink spectral efficiency is a function of large-scale static coding. A unique aspect of Massive MIMO systems is that compared to single antenna systems, these systems can support power allocation problems with great simplicity.

In order to optimize channel capacity, the minimum channel capacity must be maximized. Therefore, the problem of finding the optimal power allocation for this can be defined as the Max-min problem. The power allocation problem to maximize the minimum channel capacity of the user's (UE) downlink is obtained based on max-min based objective functions and conditions. At this time, the max-min based objective function and conditions can be expressed as Equation 11 below.

here, represents the maximum downlink transmission power.

Additionally, in order to maximize SINR and SLNR, the Max-Product problem must be optimized. In other words, the SINR optimization problem can be defined as a Max-Product problem, which is obtained based on the Max-Product-based objective function and conditions. At this time, the Max-Product objective function and conditions for optimizing the SINR and SLNR of the user (UE) can be expressed as Equation 12 below.

Similar to Equation 12 above, the SLNR optimization problem (Max-Product) can be expressed as Equation 13 below.

Figure 2 shows an example of a power allocation optimization algorithm for solving the Max-Min fairness problem in Equation 11 above.

The algorithm applied in Figure 2 is a Monte Carlo algorithm. Despite the power allocation strategy applied in the present invention, in order to find the optimal power, the result must be searched and found for the optimal value (point) among all cases, so Monte Carlo technology This is used. Specifically, as shown in Figure 2, if the difference between rate upper and rate lower is more than a certain level (delta), Algorithm 1 (Figure 2) is continuously repeated to update the rate upper and rate lower. If the difference between rate upper and rate lower is below a certain level (delta), the iteration ends because the optimal power allocation point that maximizes channel capacity has been found.

Figure 3 shows an example of a power allocation optimization algorithm for solving the Max-Product problem in Equations 12 and 13 above. As shown in Figure 3, despite the power allocation strategy, all cases must be summed to calculate the optimal power, so the Monte Carlo algorithm shown in Figure 3 is used. Therefore, it requires high complexity and execution time.

Figure 4 shows a deep neural network (DNN) structure for power allocation in a multi-cell, multi-user Massive MIMO system according to the present invention.

While the conventional method requires channel state information over a large amount of time to obtain an optimal solution, the deep neural network (DNN) structure proposed in the present invention improves this problem.

The Deep Neural Network (DNN) used in the present invention is completed through a learning process. The learning process is again divided into forward propagation and back propagation. In the forward propagation process, the DNN model derives results using input data from the learning data. However, since the DNN model used has not yet been trained, errors exist. In the backpropagation process, the loss function is used to compare the resulting data with the output result and update the weight and bias of the connection between each unit in the direction of reducing this. When updating weights and biases, the adpative moment estimation (Adam) algorithm is used. Since the DNN model that has been completed through the learning process has the characteristics of a simple mapping function, it can quickly predict the optimal power allocation even if new data (location information of each user (UE)) is input. These series of processes are processed by the server (Figure 1) and the results are delivered to the base station (BS). Through this, the base station (BS) can provide appropriate services to users.

Figure 5 shows the proposed deep neural network model for power allocation according to the present invention.

As shown in Figure 5, the location information of the user (UE) is used as a DNN input parameter input to the input layer in the deep neural network model (DNN Model) according to the present invention. This is because the location information of the user (UE) represents the main characteristics of the radio channel and network interference. The output through the output layer is represented by optimal downlink power allocation.

Figure 6 shows system capacity as a function of spectral efficiency (SE) per user (UE) by transmission power when the number of users (UE) in the base station is greater than the number of antennas.

As shown in FIG. 6, it can be seen that system capacity decreases when spectral efficiency (SE: Spectral Efficiency) increases. The reason is that when the number of users (UE) in a base station is greater than the number of antennas, the problem of insufficient available DoF (degree of freedom) for leverage multiuser diversity appears in the system.

Figure 7 shows processing time as a function of number of user (UE) locations.

As shown in Figure 7, in the basic method, processing time increases as the number of users (UE) increases. This is because channel estimation is performed first when calculating power allocation, which increases processing time. On the other hand, in the deep learning multi-cell multi-user MIMO system proposed in the present invention, the processing time for calculating power allocation uses the location information of users (UEs), even if the number of moving users (UEs) increases. Because it is processed in real time, the processing time remains constant.

Figure 8 shows the Cumulative Distribution Function (CDF) as a function of spectral efficiency (SE) per user (UE) in the Max-Product downlink.

As shown in Figure 8, the results of M-MMSE precoding and the deep learning network method based on SINR have an accuracy of 99% and can very effectively support optimal power allocation prediction to obtain maximum spectral efficiency (SE). It shows that the results of M-MMSE precoding based on SINR show better results than those based on SLNR.

Figure 9 shows energy efficiency (EE) as a function of summed spectral efficiency.

As shown in Figure 9, energy efficiency (EE) increases when spectral efficiency (SE) increases. However, when the energy efficiency (EE) reaches its maximum value, you can see that the energy efficiency (EE) suddenly drops to the 0 (zero) value. This is because as the number of users (UEs) increases, interference between cells increases, which has a seriously detrimental impact on both spectral efficiency (SE) and energy efficiency (EE). In addition, because the base station has limited power, when the number of users (UE) increases, energy efficiency (EE) increases up to the limited power of the base station, but beyond the limited power range, energy efficiency (EE) increases when the number of users (UE) increases. (EE) decreases.

As such, the present invention uses a deep neural network model based on user (UE) location information to optimize power allocation in the downlink of a multi-cell, multi-user Massive MIMO system, thereby reducing computational complexity and spectral efficiency. and energy efficiency, and can support power allocation in real time.

The present invention is not limited to the above-described embodiments, and various modifications and variations can be made by those skilled in the art to which the present invention pertains, within the scope of equivalency of the technical idea of the present invention and the scope of the claims described below. Of course, this can be achieved.

Claims (10)

In a Massive MIMO (Multi input multi output) system with multi-cell and multi-user,
Through a deep neural network (DNN) model, learning is performed to optimize and allocate the downlink power of the user (UE) based on the user's (UE) location, and based on the learning results, A downlink power allocation method in a multi-cell multi-user Massive MIMO system, characterized in that optimized downlink power allocation is predicted and used for power allocation.
According to clause 1,
The deep neural network model is
Calculate Signal-to-Interference-plus-Noise Ratio (SINR) and Signal-to-Leakage-plus-Noise Ratio (SLNR) according to the location of the user (UE), and allocate power based on the calculated SINR and SLNR. A downlink power allocation method in a multi-cell multi-user Massive MIMO system, characterized by predicting the optimized power allocation of a user (UE) by determining Max-Min or Max-Product.
According to clause 2,
In the Massive MIMO system,
When there are multiple (J) base stations (BS) with multiple (M) antennas in each cell,
Downlink power allocation method in a multi-cell multi-user Massive MIMO system, wherein the SINR and SLNR of user (UE) i in cell l are calculated using the following equation.

(here, is the user channel vector of user (UE) i in cell l, is the precoding vector of user (UE) i in cell l, is the precoding vector of user (UE) k in cell l, is the amount of noise change, and the SLNR (Signal-to-Leakage-plus-Noise-Ratio) per user is (indicated as)

(here, is the user channel vector of user (UE) k in cell l)
According to clause 3,
In the multi-cell multi-user Massive MIMO system, the channel capacity of the downlink (DL) of user (UE) i in cell l for the SINR and SLNR standards is calculated using the following equation: How to allocate downlink power.

(here, is the prelog factor for the sample portion per coherence block used for downlink (DL) data)

According to clause 4,
For optimal downlink power allocation according to the location of the user (UE),
Supports optimal downlink spectrum efficiency by selecting the user's (UE) precoding vector based on M-MMSE (multicel minimum mean-squared error),
The precoding vector of the user (UE) is

(here, is the precoding vector of user (UE) i in cell l, is a fused vector for detecting the uplink signal sent by user (UE) i in cell l)
satisfying Select ,
remind is MRT (maximum ratio transmission) fusion Assuming that it is made by The multicel minimum mean-squared error (M-MMSE) fusion of

(here, is the channel estimation of user (UE) i in cell l based on MMSE at base station j, as is the transmission power of user (UE) i in cell l, is the estimated error correlation matrix for the channel between base station j and user (UE) i in cell l, is the noise variation of the uplink (UL), In cell ㅣ

represents the identity matrix)
Downlink power allocation method in a multi-cell multi-user Massive MIMO system, characterized in that calculated through.
According to claim 4 or 5,
The Max-Min objective function and conditions for optimizing the downlink power allocation of the user (UE) are obtained through the following equation. Downlink power allocation in a multi-cell multi-user Massive MIMO system method.

(here, represents the maximum downlink transmission power)
According to claim 4 or 5,
Downlink power allocation method in a multi-cell multi-user Massive MIMO system, characterized in that the Max-Product objective function and conditions for optimizing the SINR and SLNR of the user (UE) are obtained through the following equation.


According to claim 2 or 4,
A downlink power allocation method in a multi-cell multi-user Massive MIMO system, characterized by applying a Monte Carlo technique to solve the fairness problem that occurs in the determined Max-Min.
According to clause 2,
The deep neural network model is
Using learning data and result data, learning is performed to optimize and allocate downlink power through the forward propagation process and back propagation process.
In the back propagation process, the learning results derived from the forward propagation process are compared with the resulting data using a loss function to connect each unit of the deep neural network model in the direction of reducing errors. Downlink power in a multi-cell, multi-user Massive MIMO system, characterized in that learning is performed to optimize and allocate downlink power by updating weights and biases using the Adam (adpative moment estimation) algorithm. Allocation method.
According to clause 9,
The deep neural network model in which the above learning was performed is
It is mounted on the server provided in the Massive MIMO system,
When new users (UEs) are located in each cell and location information of the new user (UE) is received from the base station (BS) to which the cell belongs, the received location information of the user (UE) is used as input data. Predict optimized downlink power allocation according to the location of the user (UE),
Downlink in a multi-cell multi-user Massive MIMO system characterized by transmitting predicted power allocation information to the corresponding base station (BS) so that it can be used for power allocation of users (UEe) at the base station (BS) How to allocate power.